Cross-layer importance evaluation for neural network pruning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Filter pruning is crucial for optimizing convolutional neural networks (CNNs) by reducing memory and inference time.
• Existing heuristic pruning methods can be computationally expensive and time-consuming.
• There is a need for efficient automated methods to determine optimal sparse structures in CNNs.

Purpose of the Study:

• To propose an efficient Cross-Layer Importance Evaluation (CIE) method for automated CNN filter pruning.
• To automatically calculate proportional relationships among convolutional layers for optimal sparsity.
• To enhance the efficiency and accuracy of CNN models for deployment on edge devices.

Main Methods:

• Implemented a grid sampling approach to prune each layer individually and record model accuracy.
• Developed contribution matrices to quantify the importance of each layer to overall model accuracy.
• Utilized a binary search algorithm to find the optimal sparse structure based on a target compression ratio.

Main Results:

• Achieved significant model compression with minimal accuracy degradation across various image classification tasks.
• Reduced FLOPs by over 50% for ResNet50 with only a 0.93% drop in top-1 accuracy.
• Compressed VGG19 parameters by 27.23× with a 2.46× throughput increase, incurring only 0.24% accuracy loss.

Conclusions:

• The proposed CIE method offers superior compression performance compared to existing pruning algorithms.
• CIE demonstrates a low time cost, making it practical for efficient CNN optimization.
• The method effectively balances efficiency and accuracy, facilitating CNN deployment on edge devices.